agenthropic

Documentation-corpus audit — 2026-07-06

Status: Findings and recommendations only. Nothing in this file changes the plan of record, approves a decision, or authorizes code. Gate A (Ivan’s approval of CD-1…CD-10 + LB1/LB2 in TODO.md) is still unsigned as of this writing.

Audience: a future agent session — any model, including Opus-class — picking up this project with zero conversational context. Everything needed to understand and act on the findings is spelled out here. Read §0 first; it defines every term and rule you need.

Produced by: the 2026-07-06 analysis session (Claude Fable 5). Method in §3.

Read this before acting on §0 (amendment, 2026-08-15). This audit was written for a context-free session and its §0 doubles as an operating brief, so two of its rules have to be corrected here rather than left to be discovered. The code exists. Implementation began on 2026-07-11 under an explicit owner override of CD-8, so §0.1’s “only documentation — zero application code” and §0.3’s rule 4 (“no production code, no scaffolding until Gate A is signed and WP-S7 reads GO”) describe a gate that has been opened. Gate A itself is partially signed as of 2026-07-10 — the three decision boxes are signed, the two physical acts are not — so the front-matter’s “still unsigned” is half stale and half still true. A session reading rule 4 today should not refuse to work; it should know that the override covered dispatching only and relaxed nothing else — not a security invariant, not the LABEL-ME ratification (all Phase-0 numbers remain PROVISIONAL), and not the commit rule.

Rule 2’s parenthetical is also stale: docs/ is no longer untracked. The rule it decorates — never commit or push without an explicit ask — is unchanged and still binding, as are all of §0.3’s security invariants in rule 5, which the built system implements rather than merely promises.

Everything else in §0 remains the correct orientation, and the finding registers below are preserved as written. Where a specific finding has since been overtaken, it is noted in place; where it has not, silence means it is still open. The current-truth snapshot for a fresh session lives in the project-state record, not here.

---

0. How to use this document

0.1 What this project is (three sentences)

agenthropic is a planned self-hosted, local-first dashboard for observing Claude Code agent/subagent activity on Ivan’s Mac Mini M4 — persisted subagent DAG, dollar-accurate cost, Telegram alerts. It is a greenfield build (decided over forking any of six audited rival dashboards). As of 2026-07-06 the repository contains only documentation — zero application code — by design: canonical decision CD-8 forbids production code until the Phase-0 feasibility spike returns GO.

0.2 Terminology (do not guess — these are project-specific)

Term	Meaning
CD-1…CD-10	The ten canonical decisions, resolved in docs/analysis/concept-analysis-v2.md §3.
LB1 / LB2	The two load-bearing decisions: LB1 = ingest primacy (JSONL-primary, contingent on Phase 0); LB2 = personal-first / commercial-clean identity.
WP-xx	Work package in docs/analysis/development-plan.md (75 WPs, 8 tracks: S spike · F foundation · D data · IN ingest · C cost · U realtime/UI · A alerts · X delivery).
DOC-xx	Documentation work package in docs/DOCS-PLAN.md (33 WPs, 6 tracks P/O/A/S/C/U).
Gate A	Ivan’s formal sign-off on CD-1…CD-10 + LB1/LB2. Unsigned. Until signed, the entire plan is recommendation-only.
WP-S7 / GO	The Phase-0 spike’s final GO / CONDITIONAL-GO / NO-GO verdict. Gates WP-F1 (the first scaffold WP) — i.e. gates all production code.
The probe	docs/analysis/phase0-probe.md — a 2026-07-04 read-only empirical probe of the real ~/.claude/projects corpus that pre-answered CD-1 with CONDITIONAL-GO → build, confidence 85. It de-risks but does not replace the formal spike.
Best-path memo	docs/analysis/best-path-decision.md — the strategic memo that sits above the development plan. Its §6 lists plan amendments that were never applied (finding AMEND-1…6).
The moat	Per the best-path memo: persistent cross-session DAG + dollar-cost attribution — only these two. (Older docs and the public site say four or five features; see finding LEDGER-23.)
P0 tests	Three merge-blocking release tests: (1) Σ tokens == JSONL exact; (2) double replay produces a byte-identical DB; (3) DAG rebuilt from JSONL alone. (As built, 2026-08-25: main is branch-protected on the ci check, so “merge-blocking” here means merge-blocking for anyone who is not the repository owner — enforce_admins is deliberately off for this single-maintainer repository; see the standing correction.)
events_raw	The immutable append-only ingest substrate (CD-2). Everything else (sessions, agents, edges, token_usage) is a deterministic, replayable projection.
OPCⁿ	An undefined token inherited from the vendor documents. Nobody has ever defined it. Flagged “define-or-drop” since v2 §7; still open (OPEN-9).
The five daily questions	The MVP requirement set (CD-10): what is running now · where did tokens/money go · what did session X spawn and why · what failed/stuck · what changed across sessions. Target: answerable in <30 s.

0.3 Binding rules for any session acting on this document

These come from CLAUDE.md (root + project) and are not optional:

1. Chat with Ivan is Bulgarian; everything written into the repo is English.
2. Never commit or push without an explicit ask. All of docs/ is currently untracked.
3. AI-harness files (CLAUDE.md, WORKLOG.md, .claude/, docs/ai/, *.docx) are git-excluded via .git/info/exclude and never enter commits or PRs.
4. No production code, no scaffolding (package.json, src/, workspaces) until Gate A is signed and WP-S7 reads GO. This audit does not change that.
5. Security invariants are non-negotiable and appear in every layer: loopback-only bind, no browser-driven subprocess/claude spawner, mandatory DASHBOARD_TOKEN with timingSafeEqual (fail startup if unset), same-origin realtime stream (SSE per CD-5), no SSRF, tunnel-only remote access, SQLite WAL + tested backups, token counts read from ~/.claude/projects/*.jsonl — never inferred.
6. Append a WORKLOG.md entry (English) per meaningful task.

---

1. Corpus inventory

Layer	Files	Git status	Role
Sources	4 .docx under due-diligence/ (vendor due-diligence v1 + v2; ideen-doklad base + EXPANDED)	excluded	Externally produced inputs. The two families sit on opposite sides of the digests: the vendor docx are upstream of DESIGN.md; the ideen-doklad docx are downstream (they cite DESIGN.md as input).
Digests	docs/ai/DESIGN.md · docs/independent-due-diligence.md	excluded / tracked-eligible	Design basis + the audit that overturned the vendor recommendation.
Dossier	docs/due-diligence/ — 6 modules + 6 per-project deep dives	tracked-eligible	file:line evidence for every rival grade (simple10 A−, hoangsonww B−, nirdiamant C+, cast C, disler C−, claude-code-templates C).
Analysis chain	docs/analysis/: concept-analysis (v1) → external-docs-review → concept-analysis-v2 → development-plan → best-path-decision → phase0-probe (+ animated-room-analysis, README)	tracked-eligible	The decision spine.
Trackers	TODO.md · DONE.md · WORKLOG.md (excluded)	mixed	Live state. Gate A lives at the top of TODO.md.
Public site	docs/site/ — 44 pages, 13 ADRs, 603 internal links, 0 broken	tracked-eligible	The only public surface (app itself is loopback-only).
Docs plan	docs/DOCS-PLAN.md	tracked-eligible	33-WP decomposition of the site.

2. The precedence chain — the single most important fact in this repo

When two documents disagree, resolve in this order (newest evidence wins):

1. docs/analysis/phase0-probe.md        (empirical measurements, 2026-07-04)
2. docs/analysis/best-path-decision.md  §6 (strategy memo — "sits above the plan")
3. docs/analysis/concept-analysis-v2.md (CD-1…CD-10, LB1/LB2)
4. docs/analysis/development-plan.md    (75-WP decomposition)
5. TODO.md                              (live tracker)

docs/ai/DESIGN.md is the design basis. The probe-§8 amendment order against it was executed on 2026-07-06 by a 50-agent propagation workflow (see the DONE.md entry “Propagated the four empirical CD-1 corrections”) — spawn-tool keying, layout mechanism, outbox demotion and the CD-1 verdict are now consistent corpus-wide (spot-verified by this audit: the only remaining Task mentions in DESIGN.md, WP-IN8 and docs/site/ are correction statements). The best-path §6 amendments (AMEND-1…6) were NOT part of that propagation and remain unapplied.

Why this matters: layers 4 and 5 are approximately six amendments behind layer 2 (see AMEND-1…6). A reader who takes development-plan.md or TODO.md literally will build the wrong v1.0 (alerts on the critical path, vector-DB stub included, dual SQLite driver, unslimmed spike).

3. Method

• All ten load-bearing strategy/decision documents were read in full, first-hand: DESIGN.md, concept-analysis v1 + v2, implementation-plan, development-plan, best-path-decision, phase0-probe, independent-due-diligence, DOCS-PLAN, TODO/DONE/README.
• Three parallel deep sweeps were run by subagents:
  1. docs/site digest — all 44 pages read; 19 cross-page inconsistencies catalogued (§6, C1–C19); every page checked against the seven security/design invariants (result: zero violations).
  2. Source-vs-digest audit — all four .docx extracted via textutil and diffed against the digests (Appendix A).
  3. Cross-consistency ledger — 25 analysis/dossier/tracker files cross-checked into a 33-decision ledger with staleness and orphan registers (§5, Appendix B).

4. Finding register

Each finding: statement · evidence · impact · fix. IDs are stable — cite them in future WORKLOG entries when resolving.

4.1 AMEND — decided amendments never applied to the plan of record

The best-path memo states “the plan is being amended to match (see §6)”. That amendment never happened. These six findings are the single largest consistency debt in the repo.

• AMEND-1 — v1.0 scope: no alerts. best-path §6.1 rules v1.0 = DAG + cost cockpit, alerts off the critical path. development-plan.md’s critical path still terminates in A5→A8→A9→A10 (alerts), and TODO.md Phases 5–6 still stage the full alerting stack as v1.0 work. Fix: re-annotate the critical path; mark Phases 5–6 as post-1.0.
• AMEND-2 — cut WP-A8/WP-A9 (alerts CRUD API/UI), keep A10. best-path §6.2. Both WPs still listed as deliverables with exit gates in dev-plan Phase 6 and TODO Phase 6.
• AMEND-3 — delete WP-X11 (vector-DB stub) entirely. best-path §6.3. TODO Phase 1 still schedules X11; docs/DOCS-PLAN.md §5 — written after the memo — explicitly says “WP-X11 stays”, directly contradicting it. Fix: delete from both, plus the DOCS-PLAN reconciliation note.
• AMEND-4 — single SQLite driver (better-sqlite3 only), Node 22 pinned. best-path §6.3. dev-plan WP-D2 still says “dual-driver”; CD-9’s copy-list still names hoangsonww’s “dual-sqlite driver” as a copyable artifact. Fix: amend WP-D2 text and strike the CD-9 line item.
• AMEND-5 — slim WP-S1 (drop the throwaway-hook block from the gating path), demote WP-S4 to liveness-only. best-path §6.6. TODO still runs both at full scope, and WP-S1 at full scope gates everything in wave 2+.
• AMEND-6 — add packages/core (server/web-import-free) to the scaffold. best-path §6.7. No scaffold WP mentions it; WP-F1’s layout predates the memo. Cheap now, expensive after scaffold.
• AMEND-7 — DESIGN.md amendment ordered by probe §8 — ✅ RESOLVED 2026-07-06. The probe proved: spawn tool is Agent/Workflow, never Task (0 Task blocks — a Task-keyed parser builds an empty DAG); layout is spawn-mechanism-driven (flat agent-*.jsonl for Agent, nested workflows/wf_*/ for Workflow — 85.2% of agent files nested), not CC-version-driven. A 50-agent propagation workflow applied this (plus the outbox demotion and the CD-1 verdict) across DESIGN.md, WP-IN8 and all of docs/site/ on 2026-07-06 (DONE.md entry), with an adversarial residual sweep reporting zero misses — independently spot-verified by this audit (every remaining Task hit is a correction statement). Kept in the register as the one worked example of precedence being materialized correctly; AMEND-1…6 still await the same treatment.

4.2 OPEN — decisions with no owner and no WP

• OPEN-1 — Retention TTL vs events_raw immutability. ADR-0004/0006/0009 enforce no-UPDATE/no-DELETE on events_raw by trigger + test; ADR-0012/WP-D10 mandate a retention TTL sweeper. Flagged open in architecture/data-model.md and operations/backup-restore.md §4; resolved nowhere. This is a genuine design collision that will surface as a red build in Track D. Recommended resolution (architect lens, §9.2): retention deletes projections only; events_raw ages out by segment archival (file-level detach), never row DML — preserving the triggers and the replay P0 test.
• OPEN-2 — 'unknown' missing from agents.status CHECK. The reference DDL (data-model.md, glossary.md, ADR-0006) allows ('working','waiting','completed','error'), but WP-IN12’s missing-Stop watchdog assigns 'unknown'. Self-flagged in three pages, fixed in none. Also an unanswered follow-up: the unknown→revert rule (v2 §7 q5).
• OPEN-3 — Redaction phase drift. ADR-0012 says “retention TTL + payload redaction from Phase 1”; operations/backup-restore.md §5 wires the redactor at WP-IN14 with a Phase-2 exit criterion. Pick one.
• OPEN-4 — Browser token transport (header vs cookie vs query). security/remote-access.md defers to security/model.md, which also doesn’t decide it. Security-sensitive, circular, unowned.
• OPEN-5 — Hook-POST auth mechanism. The invariant (authed) is decided; the mechanism (shared token vs socket peer creds) is question 8 in three architecture pages, while architecture/hooks.md prose reads as if already settled.
• OPEN-6 — Pricing data source (v2 §7). model_pricing is versioned and CI-gated (no-price-row fails CI) but nobody chose where prices come from or how verified_on gets refreshed.
• OPEN-7 — App port number and where it is configured (remote-access.md flags it).
• OPEN-8 — Coverage boundary: “>90%” vs “≥90%”. ADR-0009 contradicts itself (Decision says “>90%”, Acceptance says “blocks merges below 90%”, i.e. ≥90 passes). It’s the merge gate — the exact boundary matters. Pick ≥90% once, sweep all pages.
• OPEN-9 — “OPCⁿ” — define or drop (ADR-0002 left it open; used in two live docs).

Status of this register on 2026-08-15. The nine OPEN items were lifted out of this audit into open-decisions.md, which is now where their current state is maintained; read that file rather than this list when you need to act on one. Most of them have since been answered by implementation — the code picked a transport, a port, a hook-auth mechanism, a status vocabulary and a coverage threshold — but answered by implementation is not the same as decided, and none of the sign-off boxes in open-decisions.md has been ticked. Two of the items above are also wrong as stated and are corrected there: OPEN-6 rests on a verified_on column that does not exist in the shipped model_pricing table, and OPEN-8’s premise (a boundary between “>90%” and “≥90%”) was overtaken by a 100% threshold — while its “blocks merges” clause turned out to be the unexamined half, and was still false on 2026-08-15, because main was unprotected. (As built, 2026-08-25: main is branch-protected, the required status check is ci — the job id in .github/workflows/ci.yml, not the workflow’s CI display name — and force-pushes and branch deletion are refused. The clause holds for a contributor and, by deliberate choice, not for the repository owner: enforce_admins is off because agenthropic has one maintainer whose normal working mode is a direct push to main, and admin enforcement would lock the sole maintainer out of their own repository. See the standing correction.) OPEN-1 (retention) and OPEN-9 (OPCⁿ) remain open in the plain sense: undecided, unowned, and now with more surface area than they had here.

4.3 LOST — substantive source content that never reached the plan of record

From the source-vs-digest audit (full detail in Appendix A):

• LOST-1 — Vendor v2 §12: the 24-capability feature matrix (✓/~/✗ across all rivals: cache-token accounting, compaction-aware totals, token-burn/quota window, live activity pulse, transcript replay, snapshot/restore…). The closest thing to a target-feature checklist for agenthropic’s own UI. Dropped entirely.
• LOST-2 — Vendor v2 §13: adjacent tools. Notably: Claude Code emits a stable OpenTelemetry span schema with query_source = main/subagent/auxiliary — directly relevant to ingest strategy as a corroborating (or fallback) signal; plus LiteLLM gateway, claude-token-lens, Anthropic Console/Analytics API.
• LOST-3 — Vendor v2 §16 risk register, esp. “Claude Code hook-schema drift” (M/H) — a live risk for the greenfield build too (the parser reads undocumented internals across 7 observed CC versions) — and the hoangsonww CLA relicensing note.
• LOST-4 — Vendor v2 §17 cost-of-ownership sizing (developer-day estimates). The roadmap kept the phases and dropped all effort numbers.
• LOST-5 — EXPANDED §3: FR-01…FR-10 / NFR catalogue — incl. PRIV-01 (payload redaction) and PERF-01 (render 500+ events / 50+ agents). Partially absorbed into CDs; never as testable requirements with IDs.
• LOST-6 — EXPANDED §15: the 7-screen UX model + the honest-uncertainty principle (inferred edges and estimated costs must be visibly labelled; unknown model ⇒ “estimated”, never a silent zero). The project’s best UX thinking lives only in a git-excluded docx. See §9.5.
• LOST-7 — EXPANDED §7.1: the 10-scenario negative-test catalogue (SubagentStop before Start → anomaly flag; orphan → pending-reparent, no silent fake root; timing-attack attempt; huge payload; restart mid-session…). contributing/testing.md admits its 12-scenario gate does not reproduce them. Recovering them is ~30 minutes.
• LOST-8 — The “no true DAG” nuance. The quote the independent audit attributes to the vendor is not verbatim in either docx. The vendor scores simple10 “~” (partial) on “Orchestration DAG / graph” and 3-vs-5 on visualization. The audit’s core argument (the tie-break is contestable; the vendor’s own weighted model ranks simple10 first, 4.1 vs 4.0) stands — but the characterization “factual error” is stronger than the source warrants. Since the greenfield decision leans on this audit, soften the wording to “contestable judgment call” in docs/independent-due-diligence.md §0 and DESIGN §0. Related: DESIGN §0’s claim that greenfield is “reinforced, not weakened, by the independent audit” overstates — the audit recommended forking simple10, and vendor v1’s explicit “building from scratch is not justified” argument is never rebutted anywhere.

4.4 EMP — empirical gaps

• EMP-1 — Intra-workflow edge ordering is unproven. Reconstruction via journal.jsonl + promptId is item 11 of the probe’s parser gate but was never demonstrated. P0 test #3 (“DAG rebuilt from JSONL alone”) depends on it. No WP owns proving it before Phase 3 relies on it. Fix: add it to the formal spike (WP-S5 scope).
• EMP-2 — Depth-2 evidence is thin. The 100% depth-2 recovery claim rests on 6 edges. The formal spike’s paired-capture corpus (WP-S1) must include deep nesting on purpose.
• EMP-3 — The kill-condition is formally undecidable. best-path §7/§9 defined it as: kill if the 2-week baseline-friction log (using claude-code-templates --analytics) is empty AND the probe is messy. The probe half resolved favorably; the friction log was never started, scheduled, or closed out. Either run it (passive, near-zero effort) or record explicitly that the probe’s strength (confidence 85) retires it. The dangling state is the worst option.

4.5 PROC — process and hygiene

• PROC-1 — Root README.md is stale on three counts: (a) “ingests Claude Code lifecycle hooks into SQLite” — CD-1 inverted this to JSONL-primary with hooks as liveness; (b) “stack and structure are being decided” — the stack is effectively decided (Fastify + TypeBox, better-sqlite3 single driver, React/Vite/D3, pnpm, Node 22), pending only Gate-A signature; (c) no mention that a plan of record and a 44-page docs corpus exist.
• PROC-2 — Project CLAUDE.md lags the decision chain: still says “stack & repo structure are an open decision — do not scaffold” (the do-not-scaffold half is still correct; the open-decision half is not) and “same-origin check on the WebSocket” (canonical transport is SSE per CD-5/ADR-0007). Every future session reads this file first — it is actively misleading agents.
• PROC-3 — DONE.md is missing the docs-site milestone (2026-07-04: 22 pages + 13 ADRs authored by a 22-writer/22-reviewer fan-out, 44 pages / 603 links / 0 broken — recorded only in DOCS-PLAN’s status block and WORKLOG). The stated convention (“completed milestones move to DONE.md”) is broken for the largest artifact since the probe.
• PROC-4 — No LICENSE file. contributing/licensing.md asserts “the project’s own MIT posture” and applies Berne-convention logic to rivals (no LICENSE file ⇒ all-rights-reserved ⇒ clean-room only). By its own logic, agenthropic itself is currently all-rights-reserved. The LB2 commercial-clean hedge is void until a LICENSE lands.
• PROC-5 — Governance artifacts unowned. SECURITY.md, CODE_OF_CONDUCT.md, issue/PR templates are described in contributing/governance.md but exist nowhere and no WP owns creating them (“natural fit alongside WP-X6/X9” is a suggestion, not an assignment).
• PROC-6 — Two live phase vocabularies. Old scheme (DESIGN §9 / implementation-plan: P2 = Telegram, P4 = global DAG) vs new scheme (development-plan: Phase 5 = alerts). animated-room-analysis.md gates itself on the old numbering — its “not before Phase 2, ideally after Phase 4” now silently means something different. Any doc that says “Phase N” without naming the scheme is ambiguous.
• PROC-7 — Dossier files carry no supersession banners. docs/due-diligence/README.md and recommendation.md still instruct “fork simple10, graft cast analytics.ts ~50 LOC” — dead advice twice over (greenfield per best-path; cast is clean-room-only per CD-9). A reader entering the corpus through the dossier gets obsolete instructions with no pointer forward.

5. Decision ledger — status summary

Full 33-decision ledger was compiled; below, everything not simply OK. (“OK” = consistent everywhere: security posture — the single most consistent decision in the corpus — SSE transport in the analysis chain, React/Vite/D3, Fastify+TypeBox, scope/five questions, secrets handling, per-artifact licensing rule, pricing structure, gate discipline, token ground truth.)

Decision	Status	Where it stands
v1.0 contents / critical path	CONTRADICTED	best-path (no alerts) vs dev-plan/TODO (alerts in). AMEND-1/2.
Vector-DB WP-X11	CONTRADICTED ×3	best-path deletes; TODO schedules; DOCS-PLAN §5 blesses. AMEND-3.
SQLite driver	CONTRADICTED	best-path: single; WP-D2 + CD-9 copy-list: dual. AMEND-4.
WP-S1 scope / WP-S4 demotion	CONTRADICTED	best-path §6.6 vs TODO full-scope. AMEND-5.
Spawn-tool keying (Agent/Workflow)	RESOLVED 2026-07-06	Probe proved it; the 50-agent propagation carried it into DESIGN + WP-IN8 + the whole site. AMEND-7.
Monorepo layout	SUPERSEDED, unconsolidated	v1 packages/* → v2 apps/* → +packages/core (§6.7). No single doc states the final layout: apps/server, apps/web, packages/shared, packages/core, packages/test-fixtures, hooks/.
Durable outbox WP-IN11	SUPERSEDED cleanly	Probe: YAGNI-leaning, deferrable; TODO annotated correctly; dev-plan body text unamended.
Moat definition	CONTRADICTED in framing	Dossier/site: 4–5 features incl. alerting; best-path: 2 (persistent DAG + dollar cost). Public positioning and build strategy tell different stories. LEDGER-23.
Fleet status	INCONSISTENT ×3	roadmap “future decision” vs the-moat “not scheduled” vs faq “is scheduled”. Canonical: deferred until a second host exists (ADR-0002/0012).
Docs generator	INTERNALLY INCONSISTENT	DOC-P1 marked deferred, yet adr-docs-site-generator (ADR-0013) exists and content lives at docs/site/ (a path neither DOC-P1 option names).
Ingest primacy (LB1/CD-1)	OK in chain; STALE-TEXT in README	Root README still hooks-primary. PROC-1.
Transport SSE (CD-5)	OK in chain; residual WS wording	project CLAUDE.md + three site diagrams still say WebSocket/”WS/SSE”. PROC-2, C3.

6. docs/site cross-page inconsistencies (C1–C19)

All 44 pages respect the seven invariants — zero violations; two pages even strengthen them (auth on all endpoints including reads; restore-drill instance also loopback+token). The 19 real inconsistencies:

ID	Issue	Pages
C1	Fleet status in three versions	roadmap / the-moat / faq (+ “Phase 5+” citations blur deferred→scheduled)
C2	Fastify asserted as fact while stack is formally a leaning — the only style-guide violation in the corpus	architecture/hooks.md, security/model.md (vs overview, what-is, contributing/index)
C3	Residual “WebSocket/SSE” wording vs canonical SSE	what-is (diagram), overview (own diagram), threat-model (diagram+prose)
C4	12 hooks listed unhedged; everywhere else hedged 12-assumed/9-documented (SubagentStart, PermissionRequest, PostToolUseFailure unconfirmed)	guide/what-is-agenthropic.md
C5	agents.status CHECK lacks 'unknown'; WP-IN12 assigns it (= OPEN-2)	data-model, glossary, ADR-0006 vs troubleshooting
C6	TTL sweeper vs no-DELETE triggers (= OPEN-1)	ADR-0004/0006/0009 vs ADR-0012/WP-D10
C7	Redaction Phase 1 vs Phase 2 (= OPEN-3)	ADR-0012 vs backup-restore §5
C8	“>90%” vs “≥90%” coverage; ADR-0009 self-contradicts (= OPEN-8)	testing, model.md, ADR-0001/0004/0009 vs contributing/index, governance
C9	Illustrative DDL drifts across three pages (compaction column ×3 names; model vs model_id; PK types)	data-model vs cost-model vs ADR-0006
C10	hoangsonww DAG described oppositely	comparison (“persisted parent_agent_id”) vs the-moat (“never persisted”)
C11	Telegram phase: DESIGN §9 says 2, plan says 5	flagged in overview/data-model, unresolved at source
C12	Delegation-savings: DESIGN §9 Phase 4 vs plan Phase 3	cost-model, glossary
C13	Hook-POST auth settled-in-prose, open-in-question-list (= OPEN-5)	hooks.md vs its own open questions
C14	“Four non-negotiable constraints” vs nine-rule catalogue, no mapping	threat-model vs model.md
C15	“Four things it lacks” vs “five capabilities absent” in one page	what-is-agenthropic.md
C16	Project license absent while licensing posture asserted (= PROC-4)	licensing.md vs governance.md
C17	Circular deferral on token transport (= OPEN-4)	remote-access ↔ model.md
C18	App port uncommitted; tunnel examples have nothing to bind to (= OPEN-7)	remote-access, comparison
C19	Port name TokenReader vs TokenSource never picked	ADR-0008

Cheapest high-value fixes: C1, C3, C4, C8, C10, C16.

7. Holistic analysis

The project is an inverted pyramid: a large, high-quality documentation mass on top of zero code — intentional (CD-8), but producing three systemic effects:

1. The only real blockers are human, not technical. Everything funnels into two actions by Ivan: signing Gate A, and running the formal Phase-0 spike (S1, S4–S7; S2/S3 pre-answered by the probe). More documentation moves nothing.
2. Entropy grows with every document written from unamended sources — and shrinks only when precedence is materialized as edits. The 44-page site inherited the unverified hook catalog and pre-probe wording because it was authored before DESIGN.md was amended; the 2026-07-06 propagation then had to sweep ~24 documents to fix what one timely edit would have prevented. AMEND-1…6 are still accruing the same debt.
3. The public story and the strategy have split. The site markets a 4–5-feature moat including Telegram alerting; the ruling memo narrows the moat to two features and pushes alerting past 1.0. Read literally, dev-plan/TODO re-inflate the MVP.

Counterweights, equally real: the security core is remarkably coherent at every layer (zero invariant violations across ~70 files); data-as-truth (JSONL ground truth, persisted edges, tokens × dated price) runs unbroken through all layers; process hygiene (WORKLOG, git-exclusions, no attribution, no unasked commits) has been followed flawlessly; and the corpus is unusually self-aware — most defects found by this audit were already self-flagged in-page, just never resolved.

The pyramid is no longer inverted (note added 2026-08-15). Code exists — a little over 17,000 lines with 106 test files beside it — so the structural diagnosis that opens this section has expired. Two of the three systemic effects it predicted did not survive contact with the build, and one did. Effect 1 is the one that survived, and it survived exactly: the remaining blockers are still human. The spike ran and returned CONDITIONAL GO, Gate A is still only partially signed, and the items nobody but Ivan can perform — hand-labelling the LABEL-ME corpus, running the friction log, enabling branch protection — are still the whole of what is outstanding. Effect 2 has reversed direction: the code is now the amending authority, and documentation that disagrees with it is what gets corrected, which is why this note exists. Effect 3, the split between the public story and the strategy, is best judged against the site as it stands today rather than against this paragraph’s 2026-07-06 reading of it.

The counterweights held. The security invariants are implemented, not merely written down; ground truth is still read from the JSONL and never inferred; and the corpus’s habit of flagging its own defects in-page is the reason a build could be reconciled against it at all.

8. Business analysis

• Market. The vendor’s “no tool exists” thesis is false as written (claude-code-templates 28.4k★, claude-hud 26.1k★, ccusage 16.8k★ — all free). It survives only narrowed: nobody persists a per-instance DAG with long-horizon history and dollar attribution; all rivals derive edges at render time. That is a real, narrow, defensible moat — with a real cost: it rests on undocumented, churning Anthropic internals (7 CC versions in the observed corpus; the layout mechanism already changed once).
• Customer. One user: Ivan (LB2 admits this honestly). Value = the five daily questions in <30 s + real dollars at Max-20x usage + delegation savings. The best-path memo’s honest dissent (cost tracking is free elsewhere; opportunity cost vs kiko/servicenow-mcp) was never answered — the friction log was the designed answer and was never run (EMP-3).
• Scale vs capacity. 75 WPs + ≥90% coverage from commit one + 13 ADRs + 44 doc pages is team-grade process serving a solo owner. best-path §6 made the right cut (~15% of WPs, alerting off the critical path) — but the cut never landed (AMEND-1…6), so the project’s nominal scope today is still the uncut one. Applying §6 is the highest-ROI single action in the backlog.
• Commercial hedge. LB2 is cheap and well-designed (MIT-clean code, host_id in the first migration, no RBAC) — and literally void until a LICENSE file exists (PROC-4).

Two of these four have moved (note added 2026-08-15). The commercial hedge is no longer void: an MIT LICENSE sits at the repository root, closing PROC-4. The coverage figure in scale vs capacity is now 100% rather than ≥90%, which makes the paragraph’s point about team-grade process serving a solo owner stronger, not weaker. The customer paragraph is untouched by anything that has been built: the friction log was still never run, the dissent is still unanswered, and the “<30 s” figure in the value proposition has never been measured — it is an exit gate that has been neither passed nor failed, because nobody has timed it. The market paragraph’s warning about churning Anthropic internals is the one to keep watching; the parser now defends against layouts it has never seen in the wild, which is preparation, not evidence.

9. Senior-lens reviews

9.1 Business analyst

The five daily questions (CD-10) are the real functional requirements and they are good: user-phrased, measurable, with a numeric target (<30 s). The quantified acceptance criteria (hierarchy ≥95%, zero lost raw events, Σ tokens exact) are rare quality for a pre-code project. What’s missing: a formal FR/NFR layer in the plan of record (EXPANDED §3 had FR-01…10 + NFRs incl. PRIV-01/PERF-01; never transplanted — LOST-5), personas/JTBD (same), and upward traceability — no definition of how anyone will know the dashboard actually saves time post-launch. Cheapest fix: add the FR/NFR IDs as a column to the existing CD table, and let the friction log double as the baseline measurement.

9.2 Architect

The spine is right: events_raw (append-only, idempotency-keyed) + deterministic projection is event-sourcing-lite at exactly the right weight; persisted orchestration_edges with derived_from_event_id makes the moat auditable; nine named ports keep the core testable; SSE is the correct transport for a one-way stream.

Four real concerns, in pain order:

1. OPEN-1 (TTL vs immutability) must be resolved before Track D. Recommended: retention deletes projections only; events_raw ages out via segment-level archival (detach a closed period into an archive file; deletion is a file operation, not row DML) — the no-UPDATE/DELETE triggers and the replay P0 test survive intact.
2. The reference DDL lives in three diverging versions (C9). One source file, quoted not paraphrased by other pages, owned by Track D.
3. The parser is the project’s real architecture. The probe’s 11-item gate is the spec: two join schemas, two layouts (flat/nested), self-referential parent index, child-transcript token summation (parent rollup ≈0% — any “easy” implementation that reads the root transcript is silently wrong), compaction resets (63/117 sessions), concurrent same-slug sessions (92/117). It deserves its own module with a versioned payload contract and per-CC-version fixtures; the lost hook-schema-drift risk (LOST-3) re-enters DESIGN right here.
4. packages/core has no owner (AMEND-6). Add it to WP-F1 before scaffold; it is expensive to retrofit.

9.3 Developer

The stack is effectively decided — Fastify + TypeBox, better-sqlite3 (single driver), React/Vite/D3, pnpm, Node 22 — and CLAUDE.md should stop calling it open (PROC-2), because every agent session reads that file first. The final monorepo layout has never been consolidated into one sentence anywhere (see ledger row); writing that paragraph is high-value.

Practical risks from the implementer’s chair: (1) the hardest module is the JSONL parser, not the UI — 85% of agent files are nested and the 11-item gate should become literally the skeleton of the first test suite (WP-S5); (2) one-WP-one-PR across 75 WPs is heavy for a solo owner — after applying §6, merge mechanical WPs into wave-sized PR series (e.g. D5–D8); (3) coverage-from-commit-one means WP-F3 (the harness) must genuinely land first or early PRs lie; pick ≥90% vs >90% once (OPEN-8); (4) two site pages already promise concrete paths (apps/server/src/security/token-guard.ts) — at scaffold time, either honor them or fix the pages.

9.4 QA

On paper the strategy is above standard: three P0 release blockers, a three-tier golden corpus (raw/redacted/manifested) with four pathologies, QA stop-the-release authority, tested restore per release candidate. The gaps:

• The 12-scenario negative catalogue is not fully specified — the 10 externally-sourced scenarios (LOST-7) exist only in a git-excluded docx. Recover them.
• P0 test #3 stands on an unproven premise (EMP-1). QA should demand the intra-workflow proof as a Phase-0 addition, not discover it in Phase 3.
• The fixture redaction rule is undefined — the golden corpus will be built from Ivan’s real sessions (real paths, names, potentially secrets) with a public repo on the horizon. Privacy gap.
• Keep the Ivan-in-the-loop gates (S1 labeling, S5 tree sign-off) — the cheapest ground-truth mechanism in the plan; do not optimize them away while slimming S1 (AMEND-5).

9.5 UX/UI designer — the lens that had never been run

Measurably the emptiest zone of the corpus: across ~70 files there are zero wireframes, zero user flows, zero information architecture, zero visual language. Everything that exists on the topic: the five questions + <30 s target (CD-10), the usage/dashboard.md stub, four view names (WP-U6…U9) — and, only in the git-excluded EXPANDED docx, the 7-screen UX model (§15), the honest-uncertainty principle, and PERF-01 (LOST-6). The project’s best UX thinking is buried in a file agents cannot see.

What is concretely missing, by weight:

1. A visual language for uncertainty and status. Uniquely among rivals, this product ingests from two reconciled sources: edges can be inferred, costs estimated, status unknown (watchdog). “Inferred/estimated is always visible, never silent” is a design system, not a footnote — dashed edges? badges? confidence chips? The answer feeds back into the schema (a confidence field?), so it belongs before Phase 4, not inside it.
2. Scale reality. The real corpus: 117 sessions, ~849 nested agents — and 84/117 sessions have zero subagents. So the DAG view is mostly sparse/empty (design the empty and sparse states, not just the dense showcase), while workflow sessions need hundreds of nodes (virtualization, collapse policy, initial zoom). Nobody has decided what first-paint shows.
3. The five questions have no screens. “<30 s to understand” is a measurable target with no defined path — which question, which screen, how many clicks, what is above the fold. One page of flows per question turns the Phase-4 exit gate from a wish into a checklist.
4. Live semantics. SSE means a moving screen: what animates, what stays still, how jitter is avoided with 10 concurrent agents. The copied simple10 layoutTree/physics is a starting point; nobody recorded which of its interactions survive.
5. Smaller but real: first-run/empty experience (zero sessions after install); dark mode (terminal-native user — presumably dark default, stated nowhere); table accessibility (contrast/keyboard); money display conventions ($0.0043 vs <$0.01).

Recommendation: one lightweight design WP (WP-UX0), dependency-ordered before WP-U5: IA map, five question-to-screen flows, sketch/ASCII wireframes for the four views, the uncertainty visual language, and transplanting EXPANDED §15 + PERF-01 into the plan of record. Roughly one day of work; without it Phase 4 starts with invisible design debt and a <30 s target no screen was designed to hit.

10. Priority actions (ranked by value ÷ effort)

Items marked (Ivan) require his decision; the rest are executable by an agent session on request. Never commit any of this without an explicit ask.

1. (Ivan) Sign or explicitly defer Gate A (TODO.md top). Everything else is recommendation-only until then. Done when: the two Gate-A checkboxes flip or a dated deferral note is added.
2. Apply best-path §6 to development-plan.md + TODO.md — AMEND-1…6 in one editorial pass. Done when: no alerts on the critical path; A8/A9 cut; X11 deleted (incl. DOCS-PLAN §5 note); WP-D2 single-driver; S1 slimmed/S4 demoted; packages/core in WP-F1; each edit cites best-path §6.
3. Amend docs/ai/DESIGN.md per probe §8 (AMEND-7) — ✅ done 2026-07-06 by the propagation workflow, spot-verified by this audit. Residual from the same family still open: the WS→SSE wording sweep (C3, PROC-2) — see action 7.
4. Resolve the three schema collisions before Track D (OPEN-1/2/3): TTL-vs-immutability (segment archival recommended), add 'unknown' to the status CHECK + define the revert rule, pick redaction Phase 1 or 2. Done when: one dated decision note lands in concept-analysis-v2 §7 (or a new ADR) and data-model.md/ADR-0006 are updated.
5. Run or retire the friction log (EMP-3) (Ivan). Done when: either claude-code-templates --analytics baseline logging starts with an end date, or a dated retirement note is added to best-path §9.
6. Add LICENSE (MIT) (Ivan approves) + record the docs-site milestone in DONE.md (PROC-3/4). Done when: LICENSE exists and DONE.md has the 2026-07-04 docs entry.
7. Refresh README.md and project CLAUDE.md (PROC-1/2): JSONL-primary framing, stack-decided-pending-Gate-A, SSE. Done when: no hooks-primary or WebSocket or stack-open wording remains.
8. Recover the lost source material (LOST-1/2/3/5/6/7): feature matrix as a UI checklist appendix, OTel/adjacent-tools note, hook-schema-drift risk into DESIGN, FR/NFR IDs into the CD table, EXPANDED §15 + honest-uncertainty + PERF-01 into the plan, the 10 negative scenarios into contributing/testing.md.
9. Create WP-UX0 (§9.5) with a dependency edge before WP-U5.
10. Close the small opens: define-or-drop OPCⁿ (OPEN-9), pricing source (OPEN-6), token transport (OPEN-4), coverage boundary sweep (OPEN-8), soften the “no true DAG” wording (LOST-8), add supersession banners to the two dossier files (PROC-7).

---

Appendix A — source-vs-digest audit (the four .docx)

Causal position: the vendor due-diligence docx (v1, v2) are upstream of the digests; the ideen-doklad docx (base, EXPANDED) are downstream — they cite DESIGN.md, independent-due-diligence.md and WORKLOG as inputs. So vendor content absent from the digests is “lost in digestion”; ideen-doklad content absent is “never folded back”.

Vendor v1 → v2: same recommendation (adopt hoangsonww), but v2 introduces the 8-criterion weighted scoring model in which simple10 wins 4.1 vs 4.0, then overrides the result on the visualization axis — the exact paragraph the independent audit attacks. Grade change disler B+→B; hoangsonww test files 67→65; adds §12 feature matrix, §13 adjacent tools, §16 risk register, §17 cost model, §18 threat model, §19 integration blueprint, §20 roadmap (~6 developer-days), appendices A–G.

Ideen-doklad base → EXPANDED: 11 → 26 sections. EXPANDED adds FR/NFR catalogue, personas/JTBD, numeric MVP metrics (≥95% hierarchy, 0 lost events, <30 s), ADR-001…010, negative-test catalogue, QA gates A–F, UX screen model + honest-uncertainty, risk register R-01…R-10 (adds R-09 DB growth, R-10 false confidence from inferred data), 23-item backlog, traceability matrix, release checklist. Unresolved schema fork between the two: base = single events + webhook_targets/webhook_deliveries; EXPANDED = events_raw+events + alert_deliveries. v2/CD-4 adopted the EXPANDED shape; the alert-tables divergence resurfaces in WP-A2 and was never explicitly reconciled.

Digest claims not supported by sources: the “no true DAG” verbatim quote (LOST-8); DESIGN §0’s “reinforced, not weakened” framing (both sources recommend fork-not-build; v1 explicitly says building from scratch is not justified — never rebutted); DESIGN’s Phase 1.5 animated-room material comes from a later separate analysis, outside DESIGN’s declared sources. Also noted: the independent audit reports simple10 at 78 test files vs the vendor’s 76 without comment.

Generation-quality note: EXPANDED §§10.1–10.7 repeat an identical four-row table seven times; the base doc contains a stray Hindi-script token. Both docx are partially templated — treat their apparatus (IDs, catalogues) as valuable and their prose volume as inflated.

Appendix B — orphan register (analysis conclusions tracked nowhere)

1. best-path §6.1/§6.2/§6.3/§6.6/§6.7 edits (= AMEND-1…6).
2. best-path §6.9 — tokenless public demo with synthetic fixtures, post-1.0. No backlog item.
3. best-path §7/§9 — the friction log / kill-condition (= EMP-3).
4. probe §8 — DESIGN.md + reconstructor-spec amendment (= AMEND-7; closed 2026-07-06 by the propagation workflow — the only orphan in this register that has been adopted).
5. probe §2 — intra-workflow ordering proof (= EMP-1).
6. nirdiamant non-destructive run-checkpoint pattern (git stash + tag) — praised in the dossier as portable to agenthropic’s session model; absent from every plan.
7. v2 §7 residue — retention TTL number, coverage scope, pricing source, hook-POST auth detail, OPCⁿ (= OPEN-3/6/5/9 + coverage scope).
8. recommendation.md’s three “open questions for Ivan” — two answered implicitly by later docs, OPCⁿ never; no closure recorded.
9. DOCS-PLAN deferred WPs (DOC-P1/P2/P4/P5, Track U fills) — invisible from TODO.md, which claims to be “the only actionable work” list.
10. Animated-room approval decision — deliberately parked, but with no “decide at gate X” entry anywhere; the parking brake has no release lever.

This site is open source. Improve this page.